Perspective · for graduating students
Every link in this guide was opened and checked in August 2026. Almost all of it costs nothing. The hard part is no longer finding an AI course. It is knowing which handful actually change what a hiring manager thinks of you, and in what order to take them.
145
verified courses and resources
112
completely free, no trial
24
free to audit
5
phases from zero to hireable
Read this first
You do not need a computer science degree to work in AI in 2026. You do need a route that plays to what you already have, a sequence that is deliberately slow at the start and fast at the end, and a ring of skills around the model that courses about models never teach. This guide holds all three, plus the library to start from: the core machine learning (ML) curriculum, the large language model (LLM) work that 2026 hiring actually asks for, and the complementary skills that separate a student who has watched lectures from one who has shipped something.
“Which handful of courses actually change what a hiring manager thinks of me, and in what order?”
Start here
Find yourself below, then follow that column through the sequence.
Computer science, IT and software engineering students
Target roles: AI engineer, ML engineer, LLM application engineer, MLOps engineer.
Engineering, maths, physics, statistics, economics
Target roles: Data scientist, applied scientist, quantitative analyst, domain AI specialist.
Commerce, arts, law, design, life sciences
Target roles: AI product manager, AI governance analyst, business analyst, domain AI lead, solutions consultant.
The sequence
Deliberately slower at the start and faster at the end. Most students fail by beginning with a deep learning specialisation before they can write a clean function or a join.
Weeks 0 to 4
Two goals only: be able to talk about AI accurately, and be able to write Python without looking things up constantly.
Take: Elements of AI, then CS50P or Kaggle Python.
Non-technical students: add AI For Everyone and Generative AI for Everyone.
Proof of phase: You can explain the difference between training and inference to a friend, and you have written a script that reads a file and does something useful with it.
Months 1 to 3
Learn how models are built, and at the same time learn the four unglamorous skills every job description assumes: Git, SQL, the command line and pandas.
Take: Google ML Crash Course or the ML Specialization, plus Kaggle Intro to ML, SQLBolt, GitHub Skills and the Missing Semester.
Proof of phase: A public GitHub repo with a trained model, a README, and a query that pulls its data.
Months 3 to 6
The fork in the road. Take both branches if you can. If you have to choose, choose the branch your target role hires for.
Take: Research and modelling branch: Deep Learning Specialization, fast.ai, Hugging Face LLM Course.
Application branch: Generative AI with Large Language Models, then the DeepLearning.AI short courses on prompting, LangChain and the Claude or OpenAI APIs.
Proof of phase: A working app someone else can open in a browser.
Months 6 to 9
This is where 2026 hiring actually is. Anyone can call an API. Very few graduates can make an agent reliable, ground it in real documents, evaluate it honestly and deploy it.
Take: Hugging Face AI Agents Course, Introduction to LangGraph, a retrieval course, Evaluating AI Agents, and MLOps Zoomcamp or Made With ML. Add one cloud path from AWS, Azure or Google.
Proof of phase: A deployed retrieval or agent app with an evaluation suite and a Dockerfile.
Months 9 to 12
Pick the thing you want to be known for, and pick up the governance literacy that is now a hiring criterion rather than a nice extra.
Take: One specialisation, whether that is computer vision, natural language, agents, data engineering or a domain, plus Ethics of AI and either the NIST AI Risk Management Framework or the OWASP Top 10 for LLM Applications.
Proof of phase: Three portfolio projects, one blog post explaining a hard thing you learned, and a CV that names frameworks rather than buzzwords.
If you only have three months
Compress to this and you will still be employable, though narrower: CS50P or Kaggle Python, Google ML Crash Course, SQLBolt, GitHub Skills, ChatGPT Prompt Engineering for Developers, LangChain: Chat with Your Data, and Streamlit. Ship two projects. That is roughly ninety hours of study and it beats a half-finished specialisation every time.
The library
Search by name, provider or topic. Filter by level and by what it costs. Top picks are the ones worth doing first in their group, and they are what you see until you ask for everything.
Showing 32 of 145 · the 32 top picks
Harvard University
Functions, conditionals, loops, exceptions, libraries, unit tests, file I/O, regular expressions and object-oriented Python.
The strongest free programming foundation there is, and its rigour is what makes later ML material tractable.
Harvard University
Graph search, adversarial search, knowledge representation, probability, optimisation, neural networks and language models, all as working Python projects.
The projects are portfolio grade and the Harvard name carries real weight with recruiters.
Google for Developers
Linear and logistic regression, classification, data handling, neural networks, embeddings, language models, production systems and fairness.
Rebuilt with a language-model module, so it is the fastest structured on-ramp to applied ML.
Stanford Online + DeepLearning.AI (Andrew Ng)
Supervised learning, neural networks with TensorFlow, decision trees, unsupervised learning, recommenders and reinforcement learning.
Still the first real ML course that hiring managers recognise by name.
DeepLearning.AI on Coursera
Building and tuning neural networks, hyperparameter optimisation, project strategy, convolutional networks and sequence models including attention.
The canonical deep learning foundation, still the assumed prerequisite for most applied AI roles.
fast.ai
Training and deploying state-of-the-art vision, language, tabular and recommender models top down with PyTorch and fastai.
The fastest route from zero to a deployed working model, which is exactly what a portfolio needs.
Andrej Karpathy
Building neural networks from scratch in code: micrograd, backprop by hand, MLPs, batch norm, WaveNet, GPT and tokenisation.
Building GPT line by line is the single most persuasive thing a student can discuss in an AI interview.
Hugging Face
Transformers, tokenizers, datasets, fine-tuning, model sharing, and advanced topics including reasoning models.
Hugging Face is the de facto industry toolchain, so this maps almost one to one onto day-one job tasks.
Vanderbilt University on Coursera
Reusable prompt patterns for writing, analysis, ideation and automation, taught as a systematic technique rather than tricks.
For a non-technical student this has the highest immediate productivity return per hour on the whole list.
DeepLearning.AI with OpenAI
Using the API for summarising, inferring, transforming and expanding text, and building a custom chatbot.
Under two hours to become genuinely useful with language model APIs. Highest return single item here.
AWS and DeepLearning.AI on Coursera
The full model lifecycle: pretraining, scaling laws, prompt engineering, fine-tuning, parameter-efficient methods and alignment.
The closest thing to a standard curriculum for the LLM engineer job title now appearing in most postings.
DeepLearning.AI
A filterable catalogue of short courses built with OpenAI, Anthropic, Meta, AWS, LangChain and LlamaIndex.
The single fastest way to show current, hands-on familiarity with the exact tools AI teams use.
Hugging Face
Agent fundamentals, the think-act-observe loop, frameworks including smolagents and LangGraph, real use cases and a benchmark assignment.
The best-known free agents credential, and it produces a leaderboard-ranked final project.
LangChain Academy
Graph construction, state management, routing logic, streaming, human-in-the-loop control, memory, assistants and deployment.
The deepest free treatment of the stateful agent framework teams actually run in production.
DeepLearning.AI with Anthropic
The Model Context Protocol standard, building and deploying MCP servers, connecting chatbots to external tools and data.
MCP became the cross-vendor standard for tool access, so it is now a named requirement in agent roles.
DeepLearning.AI with LangChain
Document loading, splitting, vector stores and embeddings, retrieval strategies, and a chatbot grounded in your own documents.
Document-grounded chat is the single most common internal AI project companies fund.
DeepLearning.AI with Arize AI
Adding observability and tracing to agents, evaluating with code-based evaluators and model-as-judge, and running improvement experiments.
Agent evaluation and observability is the number one gap teams hire for once a prototype exists.
Streamlit
Building a chat app with Streamlit chat elements, a LangChain-powered app, user feedback capture, and response validation.
The fastest path from a notebook to a shareable demo a hiring manager can actually click.
Hugging Face
Hosting demo apps built with Gradio, Streamlit, Docker or static HTML on a public profile, with hardware upgrade options.
A Spaces profile is the closest thing to a standard portfolio format in AI hiring.
GitHub Skills
Repositories, branches, commits and pull requests, learned by doing them inside a real repository.
Every AI team ships through pull requests, so Git fluency is table stakes on day one.
Docker
Containers, images, registries, Dockerfiles, layers, port publishing, volumes, bind mounts and Docker Compose.
Models ship as containers, so a candidate who can write a working Dockerfile is immediately deployable.
SQLBolt
SELECT queries, filtering, sorting, joins, NULL handling, aggregates, query execution order and schema changes.
SQL is still the single most-tested practical skill in data and ML screening rounds.
DataTalks.Club
Docker, Terraform, cloud infrastructure, workflow orchestration, BigQuery warehousing, dbt, Spark and Kafka, ending in a capstone.
Produces a public end-to-end pipeline project, which is the portfolio artefact hiring managers actually open.
DataTalks.Club
Experiment tracking with MLflow, orchestration, deployment as web service, streaming and batch, monitoring, CI/CD and Terraform.
The closest free course there is to an actual junior MLOps job description.
Duke University on Coursera
Managing machine learning product development end to end, the data science process, and human-centred AI design with privacy built in.
AI product management is one of the highest-paid routes into AI for people who do not write production code.
Google on Coursera
Spreadsheets, SQL, Python, Tableau and the full analytics workflow from cleaning data to presenting findings.
Data analyst is the most common non-CS entry point into AI teams, and this is the most recognised entry credential.
Wharton, University of Pennsylvania on Coursera
Building trust, cooperation versus competition, detecting deception, active listening and effective apology.
Communication is the biggest differentiator between technically competent juniors and people who get promoted.
University of Helsinki
Non-maleficence, accountability, transparency, human rights and fairness, applied through practical AI ethics scenarios.
The best-known free university AI ethics course and a credible signal for responsible AI and policy roles.
NIST, US Department of Commerce
The Govern, Map, Measure and Manage functions for identifying and managing AI risk and building trustworthy AI.
The de facto reference framework in US enterprise and federal AI governance job descriptions.
Future of Life Institute
Navigate and search the Act by article, chapter and recital, run a compliance check, and read plain-language summaries.
Fastest way to become the person on the team who can answer whether the Act applies to a product.
OWASP GenAI Security Project
The ten highest-impact language model application risks from prompt injection to unbounded consumption, with mitigations for each.
The industry-standard checklist cited in AI security job descriptions and security reviews.
Lakera
Hands-on prompt injection across eight levels, showing why naive guardrails fail and how layered defences behave under attack.
Gives you a concrete story about breaking a model's defences, which AI security interviews reward.
The honest answer to your second question
Courses about models are now abundant and therefore cheap as a signal. What is scarce is the ring of skills around the model. Each of the six below is the reason a candidate gets picked over an equally clever one.
1
Most of the work on any AI team is getting to trustworthy data. Students arrive fluent in pandas on a clean file and helpless against a warehouse with eleven joins and three definitions of “active user”.
Start with SQLBolt and Kaggle Advanced SQL, then the Data Engineering Zoomcamp.
2
Git, the command line, Docker, testing and code review. This is what turns a notebook into something a team can own. It is also the fastest thing on this page to learn, which makes skipping it indefensible.
Start with GitHub Skills, the Missing Semester and Docker Get Started.
3
The gap between students who can train a model and students who get hired is almost always deployment. Experiment tracking, monitoring, drift, CI/CD and cost. A model that nobody can call is a hobby.
Start with MLOps Zoomcamp or Made With ML, plus one cloud path.
4
The single most in-demand skill of 2026, and the least taught. Anyone can build a demo that works most of the time. Companies pay for the person who can measure how often it fails, and why, and then close the gap.
Start with Evaluating AI Agents and Building and Evaluating Advanced RAG.
5
With the EU AI Act in force and ISO/IEC 42001 turning up in procurement questionnaires, someone in every team has to answer risk questions. Right now that person is usually improvising. Being the graduate who can name the control is a genuine edge.
Start with Ethics of AI, the NIST AI Risk Management Framework and the OWASP Top 10 for LLM Applications.
6
Two things no course teaches directly. Learn to write a clear one-page explanation of what you built and why it matters. Then pick one industry and learn its vocabulary properly. Generic AI skill plus real domain knowledge is a much rarer combination than either alone.
Start with Improving Communication Skills and one domain course from the library.
What actually gets you hired
Nobody has ever been hired for a screenshot of a completion badge. Build these five things over the year and you will have something to talk about in every interview.
One clean classical ML project
A real dataset you chose yourself, an honest evaluation, and a README that explains what you tried and what failed. Reviewers read the README first. Host the code on GitHub.
One deployed language model app
A retrieval assistant over documents you care about, whether that is your course notes, a public dataset or your college’s handbook. Deploy it on Hugging Face Spaces or Streamlit Community Cloud so the link works from a phone.
One agent that does a real task end to end
Something with tools, a failure mode you have handled, and a written evaluation of how often it succeeds. This is the project most graduates do not have.
One contribution to somebody else’s code
A documentation fix, a bug, a test. Open source contribution history is a credibility signal you cannot fake, and the first one is much easier than you expect.
One written explanation
A short post explaining one hard idea you understood properly, such as why attention works or why your first retrieval system returned nonsense. Writing is the skill that compounds fastest across an entire career.
A note on certificates. Free audit gives you the whole learning experience and no certificate. That is the right trade for almost every student. Pay for a certificate only when a specific employer or a scholarship application asks for one, or when the graded assignments are the point, as with the deep learning specialisations. Coursera offers financial aid on most individual courses, which usually takes about two weeks to be approved.
Coursera’s free audit has narrowed. Multi-course specialisations, including the Deep Learning Specialization, the NLP Specialization, Python for Everybody, IBM AI Engineering and Google AI Essentials, are now a seven-day trial followed by a subscription. Standalone single courses, such as AI For Everyone, Prompt Engineering for ChatGPT, Generative AI with Large Language Models, MLOps Getting Started and Ethics of AI, still audit free. Where a specialisation is now paid, the library names a free alternative next to it.
Google Cloud Skills Boost is now Google Skills. The old address redirects. All Google links here use the current one.
The Hugging Face NLP Course is now the LLM Course. Same home, twelve chapters, expanded with fine-tuning and reasoning models.
Microsoft’s AI-900 certification retired in June 2026 and its replacement has no published learning path yet. Use the free Azure AI learning path in the library instead, and do not put AI-900 on a study plan.
fast.ai Part 1 is still the 2022 edition. It remains outstanding for practical deep learning, but it predates the language model wave. Pair it with the Hugging Face LLM Course rather than treating it as current on generative AI.
Two DeepLearning.AI short courses have been retired and are replaced here by their current equivalents on agent evaluation and application accuracy.
How this guide was built. It began as a deep research report on free AI courses. Every course in that report was then re-checked against its official page, corrected where addresses had moved, and extended with the complementary skills the original report did not cover: cloud and data engineering, agents and retrieval, business and domain skills, and AI governance and security. Anything that could not be opened and confirmed was removed rather than listed on trust. Course availability, pricing and content change: confirm on the official page at the moment you enrol. Every course links directly to its official page. Prepared August 2026. Free to share with students.
Once you are in a security seat, the same idea applies to your own gaps.
Subscribe for the next essay.